Your point is well-taken, but this is also exactly why AI reliance is dangerous. Anyone who sees this should realize the precarity of relying on products that can just be locked away from you.
I’m not defending AI, but I can come up with >10 products that would absolutely cripple the company I work at if the provider suddenly says “Soz, terms of service violation”.
Vendor reliance is dangerous. That doesn’t just apply to AI. If the company in OP’s message had both Claude and Gemini they’d been okay, so the problem isn’t with AI explicitly - the problem is with reliance on services that are critical for workflows, and providers being able to change their mind at a moment’s notice.
In any case, leaving aside where the problem is, the idea that 60 employees can’t use Natural Intelligence to do their jobs means there’s something really wrong with that company…
It’s a Faustian bargain, a company gives up all of their internal IT staff and hardware and becomes completely dependent on a vendor for critical business processes. It’s like the opposite of insurance, they’re saving some money by risking a total loss of their ability to do business should the vendor pull support.
I don’t know why any company with hundreds of thousands of dollars to spend on commercial LLM APIs wouldn’t just build and self-host their own LLM fine-tuned on data relevant to their work…
It’s not that they can’t be productive. Right now at least, what AI does is amplify how much work you can do. One of my friends codes for a big company that uses state of the art Claude models and he says that the system does 80-90% of the coding grunt work and the job is more of an editor and making sure everything is correctly annotated so that humans can understand what’s happening in the code in the future. This means that work that might have taken months he can complete in a week or two.
This approach to coding is exactly what creates the problem. They will find out the hard way if they can continue to be productive when something breaks and AI is not available for whatever reason. Does anyone know how to fix it? Is the documentation sufficient to understand what the AI did?
My friend said early AI iterations were really bad at being opaque and that even now if you’re having it design the core architecture you’re going to have the problems you mentioned. But his job has basically changed to being focused mostly on being that architect. Using the metaphor of constructing a building. He used to have to do a lot of manual labor too, not just be an architect. Now he just has to tell the AI system what to build AND how. But the majority of the actual “construction” work is done by the AI system.
To continue with the analogy though, how many architects create things that an engineer takes one look at and laughs at because it’s structurally impossible (hint: a lot). Knowing the deep parts of the code and how it works becomes even more invaluable otherwise you risk Chinese building practices (quick, looks good, falls apart quickly).
My friend is a full stack programmer with over 15 years experience with one of the largest financial institutions. So he can handle what you’re talking about no problem. But what IS a huge problem is that the reason he has the requisite knowledge now is because he spent years learning best practices by doing the grunt work that’s going to disappear. So in a few years they might no longer have people with the skills to do things right and then what you’re describing will absolutely happen and build quality will go to hell. The assumption from big tech is by then the models will have improved enough it won’t matter by then.
That’s a hell of an assumption. Since we’re whipping out credentials, I’ve been in IT almost 30 years and I can tell you it’s not going to work like that.
Since we’re whipping out credentials, I’ve been in IT almost 30 years and I can tell you it’s not going to work like that.
I’m not the person you were replying to but I’ve also been in tech since 1996 and lots of things have worked just like that. All successful technology starts off barely functional and improves over time until nearly all members of it’s intended audience can successfully use it.
As an example in 1996 setting up a router was a specialty task that required training, by 2016 any moron could buy one off the shelf and have it running in an hour. As another example basic HTML was a specialty skill in 1996 but by 2003 you could do it with Microsoft Word. Smartphones are another example, they went from barely functional Windows Mobile and Blackberry devices which required ridiculous amounts of back end skill to deliver email to iPhones and Androids that any numskull can use for nearly anything at all.
My point is this; too many people are stuck on the “What use is a newborn baby?” question without realizing that the infant is growing-up at blinding speed. It’s also the first technology to carry the promise, real or not, of self-improvement when it reaches sufficient maturity. Assuming that happens all further improvement will be increasingly automatic and happen even faster.
AI isn’t going away and it’s only going to get better as time goes on.
You get it. I don’t understand the people in tech burying their heads in the sand. If the question were AGI that is definitely disputable in terms of even the viability. But plain old AI is already here. It’s not even a baby anymore.
I can see, in programming, how the current AI trend is displacing a lot of junior programmers who will not be senior programmers in 10 years due to the inability to obtain experience.
AI hasn’t come for DevOps or SysAdmins jobs either, but it’s ‘good enough’ to do help-desk/tier 1-type tasks. That limits the job pool for new IT workers and will create a future shortage of experienced workers.
I’m not worried about MY job, I’ve already accumulated the experience. It’s the new guys who are trying to get into support positions, where they are glorified knowledge base/Google searchers, who are having the hard time because AI CAN do search and summarization/RAG pretty effectively.
At least in my experience these models are pretty good now to write code based on best practices. If you ask for impractical things they will start doing ugly shortcuts or workarounds. A good eye catches these and you either rerun with a refined prompt, fix your own design or just keep telling it how you want to have it fixed.
You still gotta know how good code looks like to write it, but the models can help a lot.
I don’t doubt that it is possible to create good code when focusing on programming best practices etc. and taking the time to check the AI output thoroughly. Time however is a luxury most of the devs in those companies don’t have, because they are expected to have a 10x code output. And thats why the shit hits the fan. Bad code gets reviewed under pressure, reviewers burn out or bore out and the codebase deteriorates over time.
But we have to identify this as what it is: an internal policy failure where they abandon proven processes to maintain code quality.
I guess I’m lucky my managers have not put that pressure on me yet. I do however see developers getting sloppy and lazier so the reviews actually do take more effort and AI rarely catches all problems with a change.
Your point is well-taken, but this is also exactly why AI reliance is dangerous. Anyone who sees this should realize the precarity of relying on products that can just be locked away from you.
Like Gmail? Google drive? Slack?
I’m not defending AI, but I can come up with >10 products that would absolutely cripple the company I work at if the provider suddenly says “Soz, terms of service violation”.
Vendor reliance is dangerous. That doesn’t just apply to AI. If the company in OP’s message had both Claude and Gemini they’d been okay, so the problem isn’t with AI explicitly - the problem is with reliance on services that are critical for workflows, and providers being able to change their mind at a moment’s notice.
In any case, leaving aside where the problem is, the idea that 60 employees can’t use Natural Intelligence to do their jobs means there’s something really wrong with that company…
kind of a difference between infrastructure for daily operations and AI though? well, there should be…
1000% this.
It’s a Faustian bargain, a company gives up all of their internal IT staff and hardware and becomes completely dependent on a vendor for critical business processes. It’s like the opposite of insurance, they’re saving some money by risking a total loss of their ability to do business should the vendor pull support.
Windows 11, Onedrive, Intel Management Engine, Google accounts, …
France’s government is actively leaving Windows for Linux as you read this. I’m about to follow suit, too.
(obligatory Dune quote)
I don’t know why any company with hundreds of thousands of dollars to spend on commercial LLM APIs wouldn’t just build and self-host their own LLM fine-tuned on data relevant to their work…
Github, Jira, AWS, Azure, any of the cloud versions of absolutely anything reslly
It’s not that they can’t be productive. Right now at least, what AI does is amplify how much work you can do. One of my friends codes for a big company that uses state of the art Claude models and he says that the system does 80-90% of the coding grunt work and the job is more of an editor and making sure everything is correctly annotated so that humans can understand what’s happening in the code in the future. This means that work that might have taken months he can complete in a week or two.
This approach to coding is exactly what creates the problem. They will find out the hard way if they can continue to be productive when something breaks and AI is not available for whatever reason. Does anyone know how to fix it? Is the documentation sufficient to understand what the AI did?
This is how the Adeptus Mechanicus is born.
Good analogy. I’m gonna steal that :D
My friend said early AI iterations were really bad at being opaque and that even now if you’re having it design the core architecture you’re going to have the problems you mentioned. But his job has basically changed to being focused mostly on being that architect. Using the metaphor of constructing a building. He used to have to do a lot of manual labor too, not just be an architect. Now he just has to tell the AI system what to build AND how. But the majority of the actual “construction” work is done by the AI system.
To continue with the analogy though, how many architects create things that an engineer takes one look at and laughs at because it’s structurally impossible (hint: a lot). Knowing the deep parts of the code and how it works becomes even more invaluable otherwise you risk Chinese building practices (quick, looks good, falls apart quickly).
My friend is a full stack programmer with over 15 years experience with one of the largest financial institutions. So he can handle what you’re talking about no problem. But what IS a huge problem is that the reason he has the requisite knowledge now is because he spent years learning best practices by doing the grunt work that’s going to disappear. So in a few years they might no longer have people with the skills to do things right and then what you’re describing will absolutely happen and build quality will go to hell. The assumption from big tech is by then the models will have improved enough it won’t matter by then.
That’s a hell of an assumption. Since we’re whipping out credentials, I’ve been in IT almost 30 years and I can tell you it’s not going to work like that.
I’m not the person you were replying to but I’ve also been in tech since 1996 and lots of things have worked just like that. All successful technology starts off barely functional and improves over time until nearly all members of it’s intended audience can successfully use it.
As an example in 1996 setting up a router was a specialty task that required training, by 2016 any moron could buy one off the shelf and have it running in an hour. As another example basic HTML was a specialty skill in 1996 but by 2003 you could do it with Microsoft Word. Smartphones are another example, they went from barely functional Windows Mobile and Blackberry devices which required ridiculous amounts of back end skill to deliver email to iPhones and Androids that any numskull can use for nearly anything at all.
My point is this; too many people are stuck on the “What use is a newborn baby?” question without realizing that the infant is growing-up at blinding speed. It’s also the first technology to carry the promise, real or not, of self-improvement when it reaches sufficient maturity. Assuming that happens all further improvement will be increasingly automatic and happen even faster.
AI isn’t going away and it’s only going to get better as time goes on.
You get it. I don’t understand the people in tech burying their heads in the sand. If the question were AGI that is definitely disputable in terms of even the viability. But plain old AI is already here. It’s not even a baby anymore.
I can see, in programming, how the current AI trend is displacing a lot of junior programmers who will not be senior programmers in 10 years due to the inability to obtain experience.
AI hasn’t come for DevOps or SysAdmins jobs either, but it’s ‘good enough’ to do help-desk/tier 1-type tasks. That limits the job pool for new IT workers and will create a future shortage of experienced workers.
I’m not worried about MY job, I’ve already accumulated the experience. It’s the new guys who are trying to get into support positions, where they are glorified knowledge base/Google searchers, who are having the hard time because AI CAN do search and summarization/RAG pretty effectively.
Bingo!
Then you’re not dealing with cutting edge tech. Living in the past isn’t going to help you.
Thank you for assuming what I do or don’t do, or what I’m plugged into or not.
At least in my experience these models are pretty good now to write code based on best practices. If you ask for impractical things they will start doing ugly shortcuts or workarounds. A good eye catches these and you either rerun with a refined prompt, fix your own design or just keep telling it how you want to have it fixed.
You still gotta know how good code looks like to write it, but the models can help a lot.
I don’t doubt that it is possible to create good code when focusing on programming best practices etc. and taking the time to check the AI output thoroughly. Time however is a luxury most of the devs in those companies don’t have, because they are expected to have a 10x code output. And thats why the shit hits the fan. Bad code gets reviewed under pressure, reviewers burn out or bore out and the codebase deteriorates over time.
But we have to identify this as what it is: an internal policy failure where they abandon proven processes to maintain code quality.
I guess I’m lucky my managers have not put that pressure on me yet. I do however see developers getting sloppy and lazier so the reviews actually do take more effort and AI rarely catches all problems with a change.
This is what I’m hearing too. One thing my friend did mention was that without a nearly unlimited amount of tokens he’d run out really quickly.